Open Source RoboticsRobots, software & learning

BeyondMimic

Software

BeyondMimic motion-tracking training extension for Isaac Lab, with G1 tasks and a WandB motion registry workflow. [1]

Installation & first result →Official websiteGetting started

For
Researchers training a G1 motion-tracking policy [1]
Source rights
Named licenses are documented by part. Review licenses
Unknowns
compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

This BeyondMimic repository contains motion-tracking training code for Isaac Lab. It is a starting point for training a humanoid policy from a retargeted reference motion. Simulation and physical deployment are documented in a separate motion_tracking_controller repository; the broader project's deployment claims should not be read as tests of this training checkout. [1]

Is this for you?

Known limitations
  • Deployment code is a separate repository.
  • Motion conversion uploads to WandB and writes a fixed temporary filename.
  • GitHub marks the repository archived as of this review.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] [6] · Editorial assessment

What you need

Operating system
  • Linux x86-64 (README target)

[1]

Languages / runtimes
  • Python

[2]

Dependencies
  • Isaac Lab 2.1.0 and Isaac Sim 4.5.0 documented
  • Python >=3.10
  • WandB >=0.19 and a motion registry
  • External robot descriptions and permitted retargeted motion CSV

[1] [2]

Installation & first result

Open the upstream getting-started guide [1]

  • Prepare a permitted G1 motion CSV and robot descriptions.
  • Convert with scripts/csv_to_npz.py and verify the WandB registry replay.
  • Train Tracking-Flat-G1-v0 using that registry artifact, then evaluate the resulting WandB run.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] · Editorial assessment

Installation

Install Isaac Lab 2.1.0, keep this checkout outside IsaacLab, retrieve the documented robot descriptions and install source/whole_body_tracking editable.

[1]

Hardware, models & data

Inputs and outputs

Retargeted generalized-coordinate CSV becomes an NPZ motion artifact; training produces a tracking policy.

[1] [3]

Project notes

Replay one motion before starting training

The reviewed v1.0.0 targets Isaac Lab 2.1.0, Isaac Sim 4.5.0 and Python 3.10, with Linux x86-64 shown in the README. Install the extension outside the IsaacLab directory and obtain the separately hosted robot descriptions. The documented first input is a retargeted CSV using Unitree's generalized-coordinate convention. [1] [2] [6]

Convert one permitted motion with scripts/csv_to_npz.py, specifying --input_file, --input_fps and --output_name. The script defaults input timing to 30 fps and output timing to 50 fps; set those deliberately. Use the README's replay_npz.py registry route to check the motion before training Tracking-Flat-G1-v0. Evaluation uses the resulting WandB run path, which is distinct from its display name. [1] [3]

Conversion includes an upload

This conversion path writes an NPZ file to a fixed temporary location and uploads it to a WandB motion registry. It is not merely a local format converter. The script logs body positions, orientations and linear/angular velocities as well as joint state. Use data you are permitted to upload, configure the intended WandB organization and avoid concurrent conversions sharing that temporary file. These are source-based workflow checks; no upload or conversion was performed for this review. [3] [1]

The separate deployment project and the README-linked mjlab reproduction are useful follow-up paths, but neither proves that arbitrary checkpoints or motion files work across implementations. A workload-specific GPU-memory minimum and training cost are not established by this profile. [1]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Requirements: host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Oct 5, 2025 [6]
Latest release
Jul 24, 2026 [6]
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
v1.0.0 / cd65172032893724b445448818c34165846d847d · software repository

Source files and licenses

MIT training source; external robot and motion-data rights remain separate. Repository marked archived at review.

LayerAvailability / license
softwareavailable · MIT
Repository license text; external robot and motion assets have separate terms. [7]
documentationavailable · MIT
Repository license text; external robot and motion assets have separate terms. [7]

This profile covers training source and generated policies, not a supplied pretrained-policy release.

The scoped workflow uses user-supplied permitted retargeted CSV; linked external collections are not covered by this profile.

License declaration and external assets

The repository includes MIT permission text in LICENCE, with Isaac Lab attribution. External robot descriptions and motion datasets still need their own permission checks. GitHub marks this repository archived as of this review; that does not establish whether the documented workflow still runs. [7] [1] [6]

This review pins commit cd65172032893724b445448818c34165846d847d. It covers training source and instructions, not a supplied pretrained policy, downloaded robot assets or an independently reproduced simulator or hardware result. [6]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history